Known issues

Open bugs, gaps and limitations that are known but not yet fixed. Each entry says where it shows, how to reproduce it, and what the fix likely is. A fix removes the entry and adds a line under Fixed in CHANGELOG.md, with a regression test (an example with recorded output, a compile-fail case, or a unit test).

Fixed bugs are not listed here; CHANGELOG.md and git log have them.

Compiler#

• import a.b.{X, Y} does not parse. The parser has a form that imports items by name (self/parser.nx, parse_import: a . then {), and the checker handles it (has_name imports), but the lexer reads .{ as one token, the anonymous literal's opener, so written the natural way it fails with "expected a newline after declaration but found .{". The form is in no document (SPEC 11 has import a.b only), and impl http.Transport for T does not need it (it names the trait with its module). Fix, if the form is wanted: accept DotLBrace in parse_import, say so in SPEC 11, and add a spec case.
• A checkout of the compiler inside a package's directory builds as part of that package. nx looks for nexium.toml from the working directory up, and Nexium's repository has none at its root, so bootstrap/build.sh run in a checkout that sits inside a package (nexium-discord's first CI did this) reaches stage 1 and stops: the outer package's dependencies are "not fetched". Fix: stop the search at a repository's root (a .git), or give the compiler's own build a way to say it is no package's (a flag, or a nexium.toml for the repository); a case in the bootstrap suite.
• A WebAssembly build stops inside clang with Zig 0.16. nx build --target wasm32-wasi and artifact wasm turn on WebAssembly's exception handling for the panics' setjmp (decision 123), and with it on, Zig 0.16 compiles wasi-libc's own setjmp runtime (libc-top-half/musl/src/setjmp/wasm32/rt.c), on which its clang 21 fails with "error in backend: undefined tag symbol cannot be weak", and nx reports "C compilation failed". Native builds work with Zig 0.16, and every build works with the pinned 0.14.1, which the installer and the install script bring. Reproduce: Zig 0.16 first on the PATH and nx build examples/hello.nx --target wasm32-wasi, or the harness's ship suite. Fix: find what Zig 0.16 needs to build that runtime (runtime/nx_wasm_sjlj.c supplies it for 0.14, which leaves it out, and gives way to a libc that has it), and a CI job that builds for wasm32-wasi with the newest Zig beside the pinned one.

Runtime#

• A debug build does not fill a released String, List or Map buffer with 0xDD. SPEC 5.6 says a debug build fills freed storage with a fixed byte, so a view that outlived its storage reads garbage instead of the old contents by luck, and nx_free_bytes in runtime/nx_rt.h does that for what goes through it: ref class objects, arena chunks, the arguments cache. A container's buffer is released through nx_cont_free, which hands it to the root allocator as it is, and those buffers are what views point into. Reproduce: in --mode debug, take s[..].ptr of a String inside unsafe, let the String go out of scope, and read through the pointer: the old bytes are still there (with the cache of small freed blocks, decision 126, the first eight bytes of a small buffer hold the list's link instead). Fix: the fill in nx_cont_free as in nx_free_bytes, and a spec case that reads a released buffer through an unsafe pointer and sees 0xDD.
• nx leaks miscounts storage a thread allocated and the joiner releases, or the reverse. A spawned thread gets a copy of the context with counters of its own (nx_thread_start in runtime/nx_rt.h), and nx_thread_wait frees the record without adding the thread's counts to the joiner's, where a for parallel worker's counts are added back after the join. A buffer the thread's function grew and the joining thread releases is then a free with no allocation on the joiner's side: std.thread's spawn appends the result on the thread (task.result.append(r)) and join releases that list, and a channel's queue grows on the sender's thread and is released by ch.free() on the owner's. Reproduce: nx leaks tests/spec/s13_concurrency.nx reports 18446744073709551613 allocation(s) still live at exit, three frees more than allocations (two spawn results and the channel's queue). Fix: thread.join and thread.join_all take the context and add the thread's live_allocs, live_bytes and total_allocs to it as nx_parallel_for does, with the seed regenerated for the new signature, and a spec case joined under nx leaks.

Diagnostics#

Places where a mistake in a program's own code is reported inside the standard library, so that fixing it means reading std's source (the soft goal of SPEC 1; the roadmap's "Errors you can fix alone").

• A failed std.testing assertion points at std, not at the test. testing.expect_approx(total(xs) as f64, 7.0, 0.001) in a test fails with panic: expected 6.0 to be within 0.001 of 7.0 at <std>/testing.nx:37, the line in std that panicked; the test's own line is nowhere, so a test with several assertions does not say which one failed. The same holds for expect_err, expect_error, expect_contains, expect_lines and the snapshots. The builtin expect is right: it reports the test's line and the expression. Fix: a panic in a function that checks its caller's input reports the caller's location, as Rust's #[track_caller] does (a function attribute, or every std.testing function, whose panics take the call site as a hidden argument); a test whose failure is recorded with its own line.
• A std function's check on its arguments panics at std's line. hash.siphash("too short", "hello") stops with panic: siphash takes a 16-byte key at <std>/hash.nx:44, and the program's call is not in the message. Fix: the same caller location, for the std functions that panic about their arguments, and a case with the caller's line recorded.
• A generic std function refused for a type reports it three times. sort.is_sorted(Point, ps[..]), where Point has no Ord, gives the right error at the call ("is_sorted requires T: Ord, but Point does not implement Ord"), then two more for the same mistake inside <std>/sort.nx (the bound at 165:27, the < at 168:12), and the one that says how to fix it ("add derive(Ord) to the struct") is the last. Fix: a call that fails a generic's where bound does not instantiate the generic with that type, so its body is not checked, and the first error carries the hint; a compile-fail case expecting the hint and one error.
• A stack overflow is a bare crash. A recursion deeper than the stack ends the program with the system's fault (a segmentation fault, exit 139, on Linux; 0xC00000FD on Windows) and no message naming the overflow or the function that recursed. The fuzzer took one for a memory error: a mutant of tests/spec/s3_entry_stack.nx that lost its artifact cli { stack } overflowed (its mutants are no longer run). Fix: the runtime catches the overflow on an alternate stack (sigaltstack with a SIGSEGV handler on Linux and macOS, a vectored handler for EXCEPTION_STACK_OVERFLOW on Windows) and reports panic: stack overflow with the function, as a panic does; a spec case that overflows on purpose.

Self-hosting#

• nx tir --sigs omits body-dependent facts (error ids, alias types, and trait default methods materialized by calls) rather than being order-independent by construction. The full mode is complete; --sigs exists only as the first milestone of self/check.nx.

Tools and editors#

• nx fmt collapses aligned trailing comments. A struct whose fields carry comments aligned in one column (kind: u32 // 0 playing) is rewritten with a single space before each comment, losing the alignment the author chose; found on statusmith's Activity and Point of Origin's Level. Fix: when consecutive lines end in a comment, keep the column of the first (or the widest code) for the run; a tree-wide --check guards it.
• nx fmt glues a public generic struct's brace. pub struct Deque(T){ keeps no space before {, while struct Input(T) { (not pub) gets one: the rule that tells a generic literal (Pair(i32){) from a declaration finds struct before the name only when it begins the line. std.deque, std.heap and std.http's Streaming and Client are written the glued way because of it. Fix: look past pub (and the other declaration keywords) to struct/enum before the name, as the identifier rule does with declared, and respace the tree.
• nx refcounts lists a null written to an optional ref class field as a retain. refcount_walk in self/tools.nx reports every Retained node as "retain (copy of a reference)", and the checker wraps a null literal stored into a ?Node field in one, though the copy it compiles to (nx_clone_opt_Node on an empty optional) retains nothing. Reproduce: nx refcounts bench/trees.nx lists two retains on the leaf literal Node{ .left = null, .right = null } (line 9), where the emitted C has none; the report then overstates a program's reference traffic. Fix: skip a Retained whose operand is a null literal, or name it as a copy of null, and a case with the report's lines recorded.
• Formatter bar classification has no unit test. nx fmt tells closure bars from bit-or per line (self/fmt.nx, bar_role); the tree-wide --check in CI is the only guard. Add cases for |x| x | 1, a | b, f(|x| x), Task(T, R)|.

Tests and CI#

• The Windows portable zip's manual and completion scripts are written as one line. The release workflow writes nx man and the four completion scripts through a PowerShell pipeline with Set-Content -NoNewline (.github/workflows/release.yml, the portable-zip step), which splits a program's output into lines and joins them with nothing, the mistake bootstrap\build.ps1 had until it was fixed. A shell completion script on one line does not load, and nx.1 is unreadable. Reproduce: unzip the windows-msvc.zip of a release and count the lines of completions/nx.bash. Fix: write those files through cmd /c "... > file", as build.ps1 does, and a check in the release workflow that the files have more than one line.
• Suites that build files must pass --out-dir. Two cases compiling the same source into nx-out/ at once fail on Windows (the second write hits a mapped file). The harness (tests/run.nx) gives every case its own directory under nx-out/cases/; a new suite has to do the same by hand, nothing checks it.